COMBINED APPROACH FOR SENTIMENT ANALYSIS IN SLOVAK USING A DICTIONARY ANNOTATED BY PARTICLE SWARM OPTIMIZATION

Abstract

Sentiment analysis in the minor languages, such as Slovak, using dictionary approach is a difficult task. It requires a lot of human effort and it is time-consuming to prepare a reliable source of information, especially good dictionary. We propose an approach which uses a biologically inspired algorithm to find optimal polarity values for sentimental words. It applies a swarm intelligence algorithms, standard Particle Swarm Optimization (PSO) and Bare-bones Particle Swarm Optimization (BBPSO), to replace a human annotator at the moment of dictionary creation. We created two dictionaries, which were annotated by the human annotator, PSO and BBPSO. These dictionaries were compared with the result that the versions annotated by PSO and BBPSO outperformed a human annotator. Then a combined approach was used to classify reviews that do not contain words from the dictionary. These reviews decrease the classification performance significantly. The combined approach implements machine learning method to build a model based on the reviews classified by the dictionary approach. The combined approach finally reduced a number of unclassified reviews from 18% and 40.2% to 0.3% and increased the macro-F1 measure from 0.694 and 0.495 to 0.865 and 0.841.

Keywords

• sentiment analysis
• dictionary approach
• automatic dictionary annotation
• particle swarm optimization